The CSAT drop nobody puts in the business case

Most AI business cases account for savings but overlook a key cost: customer satisfaction. This is often omitted because savings are immediate, while satisfaction costs emerge later. If customer satisfaction is not included in the plan, projected savings may be overstated from the outset.

The cost benefits are clear; AI resolutions cost significantly less than human ones. However, if customer satisfaction is not considered, unexpected costs may arise months after launch, potentially after headcount decisions have been made.

What the numbers say

Aissist, an AI support software provider, estimates that AI-handled conversations score 5 to 10 CSAT points lower than those handled by humans, with an average industry CSAT of about 78 out of 100. Since Aissist does not publish supporting data, treat these figures as assumptions to validate, not benchmarks.

The key point remains: faster channels are not always better rated. MetricNet’s 2023 IT service desk benchmarks show chat CSAT at 77.5% versus 81.1% for phone. While this compares channels rather than AI versus humans, the trend is similar.

Customers are cautious from the outset. In a Gartner survey of 5,728 customers, 64% preferred companies not use AI in customer service, and 53% would consider switching to a competitor if they learned a company planned to implement it.

The savings are tangible: AI can resolve suitable tickets at a much lower cost. Aissist estimates a human-handled SaaS ticket costs $18 to $35, compared to $1 to $3 for AI, or about $5 including connectors, engineering, and platform fees. Aissist cites several industry cost guides for human figures but does not provide links. MetricNet’s independent benchmark is lower, at approximately $21 per agent-handled IT service desk contact in 2023.

Aissist claims that 92% of businesses report higher overall CSAT after deploying AI, though no source is provided and I could not verify this. If accurate, this supports the earlier point: instant answers to routine questions can raise the average, even if AI-handled conversations still score lower than those handled by people. A blended CSAT figure can obscure this difference.

Another consideration: vendors’ published AI CSAT figures are typically based on case studies involving simple, high-volume questions. Aissist acknowledges these results are often optimistic. Your own data is what matters most.

Where the drop comes from

The decline in satisfaction is not uniform; it is concentrated in specific areas, most of which can be addressed.

The largest decline occurs during handover. When AI transfers a customer to a human agent without providing conversation history, the customer must repeat information. Aissist found that poor handovers significantly reduce CSAT, even if the AI’s responses were accurate. Zendesk’s 2026 research found 74% of consumers are frustrated when required to repeat information.

AI performs well on order status, billing, and password resets, but struggles with emotional, ambiguous, or complex issues. Klarna is a frequently cited example. In early 2024, its AI assistant managed between two-thirds and three-quarters of customer chats, depending on the report. In May 2025, its chief executive told Bloomberg that “there will always be a human if you want” and that investing in human support quality was “the way of the future.” Aissist attributes this shift to declining CSAT on complex and emotional tickets. Some reports describe this as an addition to Klarna’s AI rather than a reversal, though I could not confirm either in Klarna’s own statements.

A human agent who is uncertain will typically admit it or consult a colleague, while AI may provide incorrect answers with equal confidence as correct ones. Customers notice and remember this.

How to put it in the business case

Begin by establishing your own baseline. Before launch, record CSAT for each question type your team handles, as industry averages may not be relevant. The gap will depend on your specific ticket mix.

Plan for a potential decline. Use Aissist’s estimate of 5 to 10 points lower for AI-handled conversations as a planning assumption and document it in your business case. This keeps the issue visible. If results are better, that is positive. If not, you are prepared to address it.

Be cautious with headcount reductions. By late 2025, only 20% of service leaders had reduced agent numbers due to AI, and Gartner predicts that half of the organisations planning significant cuts will abandon those plans by 2027.

Assign a cost to lower satisfaction. A reduced score increases the risk of customer attrition, and your finance team can estimate the value of a lost customer. Even a rough estimate is preferable to omitting it, as it ensures the business case is comprehensive. Zendesk’s 2026 research found that 85% of CX leaders believe customers will leave brands that cannot resolve issues on first contact.

Select which tasks AI should handle. Begin with structured, high-volume questions where AI performs well, and assign emotional or complex tickets to human agents until data supports a change.

There is strong evidence supporting AI that assists agents. A study of 5,179 support agents at a Fortune 500 software company, published in the Quarterly Journal of Economics, found that AI assistance increased issues resolved per hour by 14%, and by 34% for newer agents. Customer responses were also more positive.

Establish minimum CSAT thresholds and triggers. Decide in advance what CSAT level for AI-handled conversations would prompt returning a question type to human agents, and document this decision to avoid future disputes.

Maintain an accessible route to a human agent. Ensure customers can easily reach a person and that agents receive the full conversation history when they do.

When building business cases for AI and knowledge tools, I began by calculating the time saved. I measured how much time cross-functional teams spent on redundant research and repeatedly answering the same questions from customers, partners, and colleagues. I also assessed the cost of complex questions escalated to senior leaders. These figures provided annual time and cost savings, supporting the investment. If purchasing a tool, include total cost of ownership, not just the licence. Present the expected CSAT for AI-handled conversations alongside these savings to give leaders a complete view.

Where to start

If AI is already in use, segment your CSAT report by handler: AI only, AI to human, and human only. Pay particular attention to the middle group, as this is where handover issues are most evident.

If you are still developing the business case, include a statement on the expected CSAT for AI-handled conversations and your planned response if results are lower. This keeps the main issue visible.

Put it into practice: Split your CSAT by who handled the conversation